Reference

Digital & IT

AI for strains and bioprocesses, biosecurity of data, genomic privacy, quantum computing, production scheduling — data against promises.

The cluster survived triage by the sector’s rule: workflow shells and registries stayed out; the pages that remain are where software meets biology and loses in predictable places. A fermentation model is data-starved by construction — a bioprocess does not repeat enough to learn from. The design–build–test–learn loop is limited by its slowest, noisiest step, and that is not an algorithm. A genome cannot be anonymised, and it is not only yours. Synthetic-DNA screening is a similarity search with a defeatable threshold. A biologics plant cannot be sped up: a batch has a biological duration.

Start with AI strain-design platforms: the loop-is-limited-by-measurement logic ties the cluster together.

  • Agricultural bioinformatics cloud platforms The statistical genetics behind genomic selection, the reason prediction accuracy decays every generation, and why phenotyping rather than compute is the binding constraint on crop informatics.
  • AI for clinical trial patient recruitment Eligibility criteria as temporal logic, the negation and absence problems in clinical text, base rates that punish imperfect specificity, and why a matching model trained at one health system fails at the next.
  • Biosecurity in cyberspace Sequence-of-concern screening and the split-order and redesign attacks that defeat it, the integrity rather than confidentiality threat to instrument data, and the conflict between patching and validated systems.
  • Biopharmaceutical production scheduling Fixed culture durations, sequence-dependent changeover set by cleaning validation, validated hold times that expire, and the shared utilities that are the real bottleneck in a multiproduct facility.
  • Genomic data privacy platforms Re-identification from a few dozen SNPs, surname inference and long-range familial search, and the specific cost each cryptographic answer charges: ciphertext expansion, communication rounds, and a finite privacy budget.
  • AI bioprocess optimisation as a service The statistics of learning a bioprocess from a few dozen expensive, correlated runs, why hybrid models exist, and the specific problems introduced when the model is hosted by someone other than the manufacturer.
  • AI strain design and engineering platforms Why the combinatorial design space of a metabolic pathway cannot be searched empirically, what a model can and cannot learn from a few hundred strains, and why a model trained on one chassis transfers badly to another.
  • Bio-inspired AI architectures Why event-driven computation saves energy, what surrogate-gradient training actually does about the fact that a spike is not differentiable, and the software gap that keeps neuromorphic parts in laboratories.
  • Calibration management systems Metrological traceability, why every link in the chain adds uncertainty, how a calibration interval is derived from as-found data, and why an out-of-tolerance result is retroactive.
  • Quantum computing for molecular modelling Why correlated electronic structure scales badly on classical hardware, how a molecular Hamiltonian is mapped onto qubits, the measurement-count problem that limits variational algorithms, and what the FeMoco resource estimates actually say.